Research map / Label-efficient and robust learning / Recognition and long-tail
A Generalized Unbiased Risk Estimator for Learning with Augmented Classes
Senlin Shu, Shuo He, Haobo Wang, Hongxin Wei, Tao Xiang, Lei Feng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use generalized URE with generalized cross entropy and risk-penalty regularization instead of EULAC | LAC classification · MNIST, Fashion-MNIST, Kuzushiji-MNIST, SVHN | accuracy: MNIST: 0.952±0.006 -> 0.960±0.006; Fashion: 0.840±0.034 -> 0.875±0.010; Kuzushiji: 0.881±0.016 -> 0.927±0.009; SVHN: 0.8 | Improved |
| Use generalized URE with generalized cross entropy and risk-penalty regularization instead of EULAC | LAC classification · MNIST, Fashion-MNIST, Kuzushiji-MNIST, SVHN | Macro-F1: MNIST: 0.951±0.007 -> 0.960±0.006; Fashion: 0.813±0.070 -> 0.872±0.010; Kuzushiji: 0.877±0.021 -> 0.926±0.008; SVHN: 0.8 | Improved |
| Use generalized URE with generalized cross entropy and risk-penalty regularization instead of EULAC | LAC classification · MNIST, Fashion-MNIST, Kuzushiji-MNIST, SVHN | AUC: MNIST: 0.996±0.002 -> 0.998±0.001; Fashion: 0.953±0.039 -> 0.985±0.002; Kuzushiji: 0.981±0.009 -> 0.994±0.001; SVHN: 0.9 | Improved |
| Replace risk-penalty regularization with ReLU risk correction (ablation) | NRPR · MNIST, Fashion-MNIST, Kuzushiji-MNIST, SVHN | accuracy: NRPR: 0.960±0.006 vs ReLU: 0.865±0.034; 0.870±0.009 vs 0.809±0.021; 0.927±0.009 vs 0.853±0.029; 0.865±0.010 vs 0.734±0.0 | Hurt |
| Replace risk-penalty regularization with ABS risk correction (ablation) | NRPR · MNIST, Fashion-MNIST, Kuzushiji-MNIST, SVHN | accuracy: NRPR: 0.960±0.006 vs ABS: 0.776±0.014; 0.870±0.009 vs 0.747±0.019; 0.927±0.009 vs 0.757±0.027; 0.865±0.010 vs 0.695±0.00 | Hurt |
| Add risk-penalty regularization with an intermediate λ rather than setting λ=0 (ablation) | NRPR · Four regular-scale datasets | accuracy, Macro-F1, and AUC | Improved |
| Account for known-class prior shift during training (ablation) | NRPR · Usps and Optdigits with class prior shift | accuracy | Improved |
| Increase the number of unlabeled examples (ablation) | NRPR · Four regular-scale datasets | accuracy, Macro-F1, and AUC | Improved |
| Vary the assumed mixture proportion (ablation) | NRPR · Usps and Optdigits | accuracy | Mixed |
Closest papers by meaning
- An Unbiased Risk Estimator for Learning with Augmented Classes (NeurIPS 2020 · similarity 0.95)
- Unbiased Risk Estimators Can Mislead: A Case Study of Learning with Complementary Labels (ICML 2020 · similarity 0.88)
- Positive-Unlabeled Learning with Non-Negative Risk Estimator (NeurIPS 2017 · similarity 0.88)
- Complementary-Label Learning for Arbitrary Losses and Models (ICML 2019 · similarity 0.87)
- Learning from Positive and Unlabeled Data with Arbitrary Positive Shift (NeurIPS 2020 · similarity 0.86)
- On the Provable Advantage of Unsupervised Pretraining (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
